v0.2.0

## Highlights

- **[Nemotron 3 Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization).** Support for the high-resolution Sortformer architecture with native 10 ms output, for both standalone diarization and diarized ASR. (#50, #52)
- **[MagpieTTS](https://huggingface.co/nvidia/magpie_tts_multilingual_357m) v2607.** Frame-stacking support, with Arabic, Korean and Portuguese added. The model index now points to the v2607 GGUF. (#17, #42)
- **Faster MagpieTTS on CUDA and CPU.** Persistent decoder kernels, fused NanoCodec convolutions, prefetching across text chunks, and faster CPU GEMMs. See [BENCHMARK.md](https://github.com/NVIDIA/NeMo-Speech.cpp/blob/v0.2.0/BENCHMARK.md) for measured results. (#51)
- **[Nemotron Labs VoiceChat](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B).** Full-duplex speech-to-speech with realtime audio, transcripts and tool calling, served over the realtime WebSocket API. VoiceChat is CUDA-only and is built from source with the `cuda-s2s` preset; it isn't included in the release archives. See the [VoiceChat guide](https://github.com/NVIDIA/NeMo-Speech.cpp/blob/v0.2.0/docs/s2s/README.md). (#21, #39)
- **Live diarized transcription.** `transcribe --live --diarize` shows speaker turns as you speak, and `--live` now enables endpointing by default. (#50)
- **Upgraded llama.cpp.** ggml now comes from a single pinned `llama.cpp` submodule (b11151) instead of a separate `ggml` submodule. The project's patches in `patches/` are applied automatically at configure time. (#57)

## Fixes

- MagpieTTS no longer returns silence on M5-class Macs. (#20)
- Fixed nondeterministic NanoCodec audio corruption on CUDA, caused by a fused kernel reading a buffer the graph allocator had reused. (#28)
- MagpieTTS builds and runs on stock ggml (Metal, Vulkan) without the patched fused attention. (#24, #51)
- Streaming ASR keeps acoustic and decoder context across endpoints, and late punctuation no longer leaks into the next final. (#50)
- Skinny-Q8 CUDA results no longer depend on earlier requests. (#51)
- Linux Vulkan archives use the system C++ runtime, so Mesa drivers (AMD, Intel) load on current distributions. (#58)

## Notes

- **Nightly builds.** Install the latest `main` with `--channel nightly`, rebuilt daily. (#58)
- **Text normalization.** Linux and macOS archives now include inverse text normalization (ITN) and text normalization (TN). (#58)

## New Contributors
* @danilogr made their first contribution in https://github.com/NVIDIA/NeMo-Speech.cpp/pull/20
* @ryanleary made their first contribution in https://github.com/NVIDIA/NeMo-Speech.cpp/pull/28
* @virajkarandikar made their first contribution in https://github.com/NVIDIA/NeMo-Speech.cpp/pull/39

**Full Changelog**: https://github.com/NVIDIA/NeMo-Speech.cpp/compare/v0.1.0...v0.2.0